基于深度相机的三维人体重建及在服装展示方面的技术研究

【作者】 李晓捷；

【导师】 李晓久；

【作者基本信息】 天津工业大学 ， 服装设计与工程， 2016， 博士

【摘要】 人体的三维数据模型已在游戏动画、服装CAD/CAM.医学研究等领域得到广泛应用。人体体型数据的获取与表示是其中的关键技术。近年来它也成为计算机图形学、计算机仿真学和服装CAD领域的研究热点之一。例如,随着现代服装工业和数字化服装技术发展,数字化个性化服装定制(Electronic Made to Measure,eMTM)已成为趋势。提高人体特征尺寸测量的准确性,完成人体体型自动识别与分析,最终实现人体体型与服装型号的自动匹配是eMTM的发展目标。然而,在服装个性化定制过程中,人体特征尺寸的手工测量和数据整理不仅工作量大且容易出错。因此,非接触式三维测量技术成为现代人体特征测量的主要方式。它为三维人体建模、数据挖掘及虚拟试衣的进一步发展奠定了基础。基于激光扫描的非接触式人体测量设备对人眼可能造成伤害并且价格较高不宜普及,本课题利用微软公司消费级点云采集设备——Kinect对待测人体进行扫描重建,并获取人体模型的特征尺寸；然后对获得的人体数据进行体型分类；并通过特征点匹配方法实现虚拟试穿。主要研究内容如下：1、提出了一种基于多相机测量系统。本研究采用四台Kinect设备通过平行双目视觉原理确定相机内外参数。包括相机垂直光轴位置、相机光心到基准参考面距离,光心到光心的距离的标定等。2、提出一种新的曲面拟合算法。首先,为解决利用Kinect相机采集到的人体深度图像中的噪声特点,本文提出光顺去噪法对由深度图像恢复出的人体点云数据进行平滑降噪；并针对人体点云数据中出现的不同类型孔洞提出了填充优化算法；随后,利用改进的迭代最近点(Iterative Closed Points, ICP)算法将多个相机获得的人体点云图像进行配准整合；最后,通过贪婪投影三角化算法实现表面点云的三维重建,形成人体曲面模型。3、提出一种人体特征尺寸的测量法。首先,利用改进的抛物线型函数构造出人体特征点隶属度函数以确定特征点的位置；用预匹配法扫描分层特征点云,通过改变搜索步长方法实现特征点、线、面的快速提取并得出相应的特征尺寸。4、提出一种人体体型分类和识别方法。综合利用主成分分析法和模糊核聚类分析法(FKCM)将人体体型分为6类。然后通过遗传算子-支持向量机(GA-SVM)建立人体体型识别模型,针对人体体型识别误差引入遗传算法优化SVM参数进行修正,提高识别准确率。5、提出虚拟服装与人体模型的快速匹配法。假设人体模型与服装模型皆为刚体模型。根据曲率与三角约束法求出匹配特征点；并利用最小距离目标法配准曲面,实现人体与服装的简单试穿。为减小匹配误差,引入了LM算法对三维物体空间变换进行优化。并采用局部三角网格变形调整防止局部穿透问题。更多还原

【Abstract】 The three-dimensional human body model has being widely used in game animation, clothing CAD/CAM, medical and other fields, and the model data’s acquisition and representation is the key technology in this field. In recent years, the three-dimensional human body surface modeling technique has become the research hotspots in computer graphics, computer simulation and clothing CAD fields. For example, the Electronic Made To Measure (eMTM) have became the industry trends with the development of modern clothing industry and the digital clothing technology. The eMTM’s development objectives were to improve the accuracy of anthropometric data, accomplish the recognition and analyze body types, then finally matching human body to the garment. However, the artificial method was too time wasting and error-prone to eMTM. So non-contact 3D body measurement becomes the main way, and it provides the possibility for three-dimensional modeling, data mining and the virtual dressing. Since the traditional laser non-contact anthropometric devices are harm to the eyes and high prices, in this thesis I use a new points-cloud scan sensor which is produced by Microsoft Kinect to acquire the human body data and reconstruct the surface. Then the characterized sizes of human body model are calculated. After that,3D human body data are classified in terms of somatotypes. According to points matching method, we implement the virtual dry-on. The main research contents are as follows:1. A multi-cameras measuring system is proposed.4 Kinect devices are applied in our research. Inner and external reference parameters are confirmed by the parallel binocular vision theory, which includes the vertical axis arguments, the distance between camera’s optical center to its reference surface, and the distance between optical centers and so on.2. A novel three-dimensional surface fitting method is proposed. Firstly, in order to solve the problem of noise interference in the Kinect depth images, we utilize smooth fairing method to eliminate noise. And then variable types of holes are filled by optimal algorithm. Finally an improved Iterative Closed Points (TCP) algorithm is used to match the whole images obtained from multi-cameras. We utilize greedy projection triangulation algorithm to implement three-dimensional human body surface reconstruction.3. Human body characterized sizes measured method is proposed. A modified parabolic function with construct the membership function is proposed to confirm the positions of mark points. After that, we use pre-matching method to scan hierarchical point-layers. As well according to changing the searching steps of algorithm, we complete the extraction of characterized points, lines and faces. And the characterized sizes of model are also acquired.4. A classifier method is proposed to recognize the somatotypes. Here we composite the method of principal component analysis and kernel clustering analysis to divide the human somatotypes into six types. And then we use the GA-SVM to establish the human body identifying model. The genetic algorithm is introduced to amend the SVM parameters to improve the recognition accuracy.5. A fast matching method of virtual closes and body model is proposed. We assume the closes models and human body models are rigid models. We acquire matching vertex according to the curvature and triangle constraint method. Then we fit the surfaces using minimize target distances method. A dry-on effect is demonstrated finally. In addition, to reduce the matching errors we utilize LM algorithm to optimize the three-dimensional transforms as well as use adjusted local rigid triangle to avoid the garment being penetrated by body model.更多还原